Skip to content
ChristopherOx Alpha
Index
  1. Briefing
  2. Stage Learn · 2/5

Learn

From absolute zero

22 lessons. Parameters, tokens, MoE, analogies. Each one ends by tying the idea back to Ox Alpha.

  1. You will understand
  2. One lesson at a time, with analogy and trap.
  3. Tie each idea back to Ox Alpha at the end.
  4. Read all 22 if you want the full file.

Lesson 01

What an AI model is

It is not a person. It is a pattern-completion machine.

A large language model (LLM) is a program that has seen huge amounts of text and learned to answer: “what is the most likely next piece?”

Read all

It does not “understand” like you. It packed language statistics into numbers. Give it “The sky is” and it often continues “blue”, because that happened a lot in its data — and because context pushes it.

Ox Alpha is one of those. They dress it as “reasoning” (it drafts in silence first) and train it to code for hours, not just to chat.

Analogy, example, trap

Analogy. Phone autocomplete that got addicted to studying. Your keyboard predicts “th” → “the”. This one predicts whole chapters of code.

Example. When WhatsApp suggests the next word, that is a tiny model. Ox Alpha is the same idea with a stadium of numbers instead of a pocket dictionary.

Ox cut. OpenRouter describes it as a reasoning model for coding, long agents, and production.

Lesson 02

What parameters are (“fewer parameters”)

A parameter is a knob. More knobs ≠ automatically smarter.

A parameter is a number the model tunes while training. There are billions. They are artificial synapses.

Read all

People say “it has fewer parameters” like an insult. In 2026 the opposite happened: GLM-5.3 is around 744 billion (750B) and fights models ~2–3× that size (e.g. Kimi K3). Interconnects said it plainly: ~750B, a third of Kimi, still frontier at coding.

Ox Alpha does not publish a count. If it is GLM-5.x, it inherits ~744B total and ~40B active per token. “Fewer” here means “more efficient”, not “dumber”.

Analogy, example, trap

Analogy. An orchestra of 744 musicians where only 40 play at once (MoE). It sounds like a symphony, costs like a band. A dense 2T model is 2 trillion playing ALL the time: more expensive, not always more musical.

Example. An 8 GB phone is not “worse” than a 2 TB server if you run a light app. The trick is how much capacity you USE per step, not the marketing number.

Ox cut. If you hear “Ox Alpha has few parameters”, translate: probably a GLM-style MoE, not a toy. Nobody has published the official count.

Formula

cost ≈ k · P_active · tokens   (not P_total)

You pay for experts that switch on. GLM-5: 40B active of 744B. That is how it can be “small” on the bill and “big” in knowledge.

Lesson 03

MoE: the specialist team

Not one brain. A building of experts and a doorman.

Mixture of Experts: many sub-brains. A doorman decides who wakes up for each scrap of text.

Read all

GLM-5 is a huge building (744 billion knobs total) where only about 40 billion work on each word. The rest wait.

“Flash” or “Air” is usually the sibling that wakes even fewer specialists: cheaper to serve. That is how someone can give away a weekend of tokens.

Analogy, example, trap

Analogy. A hospital: you do not wake 744 doctors for a sprain. Triage calls 40. The hospital’s knowledge is still huge.

Example. Netflix does not transcode 4K for everyone. It picks a bitrate “expert”. MoE does the same with neurons.

Ox cut. The Flash/Air thesis: Ox Alpha is the “light” multimodal GLM-5.3, cheap enough to gift 100T tokens/day.

Lesson 04

What a token is

It does not count letters. It counts Lego bricks of language.

A token is a scrap of text. “Hi” can be 1 token. “extraordinarily” can be 3. An emoji can be 1 or 2. Code splits weirdly.

Read all

1 million context tokens ≈ a fat book + a whole repo. 131,072 output ≈ a long program without cutting.

“100T tokens a day” = 100 trillion pieces served. That is a factory number, not a hobby.

Analogy, example, trap

Analogy. If text is a necklace, tokens are the beads. Different makers cut beads differently. That is why the tokenizer is a fingerprint.

Example. Old SMS counted 160 characters. WhatsApp does not. Models bill (or remember) in tokens, not letters.

Ox cut. Ox Alpha and GLM-5.2/5.3 advertise 1,048,576. That exact number is not casual.

Formula

1,048,576 = 2²⁰   (a power of two, classic engineering)

Not a pretty “one million”. Exactly 2^20. GLM-5.2 already used 1M. A design coincidence, not round marketing.

Lesson 05

What a tokenizer is (the dictionary)

Before thinking, the model cuts text with its own knife.

The tokenizer is the recipe that turns letters into numeric IDs. Each lab has its own (or copies one).

Read all

If two APIs, given the same text, report the same prompt_tokens (except a FIXED extra), they almost surely share a tokenizer.

That does NOT prove the weights are equal. It proves the dictionary is the same — and rivals do not copy dictionaries by accident. China has lookalike families (Qwen/GLM), but public tests say: Ox matches GLM-5.3 and DIVERGES from Kimi, Qwen, MiMo and MiniMax.

Analogy, example, trap

Analogy. Two people using the same 2000s T9 keypad. Different messages, identical key-press count for “hello”.

Example. iPhone and Android count “a word” differently in some languages. If two apps always count the same, they share an engine.

Ox cut. This is THE star test. A teenager can repeat it: send the same text to both APIs and subtract.

Formula

tok(Ox, P) − tok(GLM-5.3, P) = 75  (constant)

75 is the wrapper (hidden system prompt). If the tokenizer changed, 75 would become 71, 80, 90 by language. It does not.

Lesson 06

What a model fingerprint is

We cannot open the skull. We measure habits.

Fingerprinting: a battery of tests a model cannot stop “being itself”. Tokenizer, API errors, video encoder, temperature-0 style, dirty tokens, audio rejection, cache.

Read all

No single test is a passport. The set is a police sketch.

Temperature 0 = always pick the most likely token. If two systems write character-for-character the same, they are the same — or one is a near-literal copy.

Analogy, example, trap

Analogy. You never see the thief’s face. You see shoe size, accent, coffee brand and the bus route. Five matches are not a whim.

Example. Netflix “knows” you by what you pause, not by your ID. Here we know labs by how they fail.

Ox cut. This dossier stacks 7 fingerprints. Six point at Z.ai. One (the playbook) could also be Xiaomi.

Lesson 07

What a stealth model is

A test car with no badge, on a public highway.

OpenRouter lets a lab serve a model as “Stealth / Animal Alpha”. People use it free. The lab watches real quality, bugs, and whether traffic holds. Then the mask comes off… or not.

Read all

4 of 4 previous: Pony→GLM-5, Hunter→MiMo-V2-Pro, Elephant→Ling-2.6-flash, Owl→LongCat-2.0. All Chinese labs.

Ox Alpha is the fifth. The first to advertise native video. OpenRouter says: we are not the owner; the provider keeps prompts (not for training, they say). OpenCode promises “zero retention”. That contradiction matters: do not paste secrets.

Analogy, example, trap

Analogy. A new restaurant opens as “Nameless Place”. Free week. Then the sign goes up. The previous four were different Chinese kitchens.

Example. “Closed” game betas that are actually open, to see if servers survive Friday night.

Ox cut. The pattern is not a passport, but the channel has a record. Betting on a Chinese lab is 4/4 statistics, not prejudice.

Lesson 08

1 million context: the work desk

Not infinite memory. A huge desk that can still get messy.

The context window is how much text fits at once: the prompt + what it generates.

Read all

1,048,576 tokens can hold a monorepo, PDFs and a long chat. It does NOT guarantee the model uses the end and the start well (the famous “lost in the middle”).

GLM-5.2 added IndexShare so full attention is not paid at 1M (2.9× fewer indexer FLOPs). If Ox Alpha holds 1M cheaply, that engineering fits.

Analogy, example, trap

Analogy. A 1 km school desk. You can put every notebook on it. You can still lose math at meter 400.

Example. Chrome with 200 tabs “fits”. That does not mean you find the tax tab.

Ox cut. Nobody has audited that Ox Alpha truly uses the million. The number is on the sheet; quality at 800K is another exam.

Lesson 09

Video encoder: how a model “sees”

Video does not enter like Netflix. It enters as a pile of expensive tokens.

An encoder splits video into frames (or space-time tubes) and turns them into tokens. Each lab picks resolution, fps and tokens per patch.

Read all

If Ox Alpha “charges” the same tokens for the same clip as GLM-5V-Turbo, the cashier is the same.

Accepting video and rejecting audio cuts candidates: Inkling (Mira) has audio. Public GLM-5.3 has neither video. So the thesis is “unpublished GLM-5.3V / Flash multimodal”, not the website 5.3.

Analogy, example, trap

Analogy. Translating a movie into a comic. If two studios use the same panels per minute, they copied the storyboard.

Example. WhatsApp compresses your videos to a bitrate. You can tell WhatsApp from Telegram by how blocky it looks.

Ox cut. Native video is what makes Ox Alpha different from public GLM-5.3 — and what pulls it toward GLM-5V.

Lesson 10

Prefix cache (the 97%)

If you paste the same repo 40 times, a smart server does not reread it whole.

Prefix cache: the model stores compute from the start of the prompt. In coding agents the repo barely changes. High hit rate = serious infra and the same caching recipe.

Read all

One tester on pi.dev measured ~97% on agent sessions. The OpenRouter catalog lists ~87% Ox vs ~92% GLM-5.3. Useful, not unique: not a passport.

This is not the model’s brain. It is the kitchen. Kitchens are not lent out.

Analogy, example, trap

Analogy. A barista who already has your cup marked. 97 of 100 coffees come from the thermos, not the grinder.

Example. TikTok does not re-encode the same video for every replay. The CDN caches it. Here the “CDN” is the LLM KV cache.

Ox cut. Points at the Z.ai stack more than Xiaomi or the West.

Formula

H = cached_tokens / input_tokens

H ≈ 0.97 on Ox and z.ai. A serving signature, not marketing.

Lesson 11

What a benchmark is (and why 80% is not the truth)

A 10-question quiz is not the national exam.

Official DeepSWE has 113 engineering tasks. A user ran 10 and Ox Alpha scored ~80%, above Fable (~65%) and GPT-5.6-sol (~52%) on THAT slice. Not an audited leaderboard.

Read all

Kingbench: Ox 87.5%, GLM-5.3 91.25%. Siblings. GLM wins by a bit; Ox is not a lazy clone nor a different god.

Always ask: N tasks? who paid? did the test leak into training?

Analogy, example, trap

Analogy. Hit 10 free throws and claim you are better than the NBA. Maybe. Maybe you picked the easy ones.

Example. A 10-sign driving quiz vs the real road test. Ox Alpha looks strong at agent coding; no official Artificial Analysis as of Aug 21.

Ox cut. We use benches as family clues (it looks like GLM), not as a gold medal.

Lesson 12

What about Mira Murati? (Inkling is not Ox)

In Spanish “mira” means “look”. Inkling is another model, already with a passport.

Thinking Machines Lab (Mira Murati, ex-OpenAI CTO) shipped Inkling on July 15–16 2026: open-weights, multimodal, customizable, ~975B MoE / 41B active, text+image+audio, 1M context.

Read all

Ox Alpha was born August 20 on OpenRouter’s stealth channel, no public weights, with video and WITHOUT audio.

No tokenizer match, no 1210 errors, no reason for Murati to hide a model she already showed in WSJ, Axios and Silicon Republic.

Analogy, example, trap

Analogy. Mira already dropped the album. Ox Alpha is an anonymous SoundCloud leak with another voice.

Example. Confusing a Tesla Semi with a tarped truck in a Chinese port because both are big. The Semi already has wheels in view.

Ox cut. Inkling is in the dossier to close a name mix-up, not as a serious suspect (3/100).

Lesson 13

What a model chat is (the loop)

You type. The model guesses the next piece. Repeat. That is the whole trick.

There is no extra “thought” besides predicting tokens. Reasoning mode means: first emit private tokens (a draft), then the answer. Ox Alpha, like GLM-5.3, seems to keep that draft always on.

Read all

A “conversation” is a list: system (who you are) + your messages + its messages. The tokenizer +75 is almost certainly that system message OpenRouter pastes in front.

Change temperature, change how much it may improvise. At 0 it is a disciplined parrot. At 1 it is more creative and more chaotic.

Analogy, example, trap

Analogy. A friend who can only speak one word at a time, but is great at guessing the next. “Reasoning” is writing a cheat sheet first, then reading you the clean answer.

Example. When Gmail completes “See you on… Friday”, that is the same loop, miniature, no cheat sheet.

Ox cut. Testers at temperature 0 saw near character-for-character matches with GLM-5.3. That does not happen across families.

Lesson 14

What an API is (the counter)

You do not talk to the brain. You talk to a window that bills in tokens.

API = a counter with rules. You send JSON (text, sometimes image). Tokens come back. OpenRouter is a mall: one counter, many storerooms behind it.

Read all

stealth/ox-alpha is the aisle name, not the storeroom. The storeroom hides. That is why fingerprints (1210 errors, 97% cache) matter: they are the storeroom ticket.

An API error is the storeroom yelling in its own language. If it yells in Chinese with code 1210, and only z.ai uses that code, you know the neighborhood.

Analogy, example, trap

Analogy. You ordered delivery with no logo. The box, the receipt and the complaint number belong to one restaurant.

Example. When a taxi app fails and says “error 12-B”, the pretty app did not invent that number. The backend did.

Ox cut. Pritish Yuvraj (ex-Llama/Meta) saw 1210/1214 identical to chat.z.ai. That is Z.ai’s counter, not Xiaomi’s.

Lesson 15

The lab map (who is who in 2026)

Not “China vs USA”. A map of kitchens with different recipes.

Closed West: OpenAI, Anthropic, Google, xAI. Own tokenizer, own errors, almost never stealth on OpenRouter with an animal name.

Read all

Open/semi China: Z.ai (GLM), Xiaomi (MiMo), DeepSeek, Moonshot (Kimi), Alibaba (Qwen), Meituan (LongCat), Ant (Ling). OpenRouter’s 2026 stealth channel was 4/4 this group.

US open: Thinking Machines (Inkling, Mira Murati). Public, downloadable weights, audio. No need for a mask.

Analogy, example, trap

Analogy. Four nameless restaurants opened on the same street this year. All four, when the sign went up, were from the same city. The fifth… smells like that city again.

Example. Not silicon racism. The channel (OpenRouter stealth) became those labs’ test track. Like a racetrack only one country’s teams use.

Ox cut. Ox Alpha is the fifth animal. Betting on a Chinese lab is channel statistics, not a hunch.

Lesson 16

What Flash / Air / 300B means (the rumor)

Flash is not a flavor. It is “same building, fewer people awake, cheaper to serve”.

An August 2026 leak (Leo / @synthwavedd, covered by AGTP) pointed at “GLM-5.3 Flash” as Ox Alpha: lighter and faster than the Aug 14 5.3. Community guess: ~300B, unconfirmed.

Read all

It fits gifting 100T tokens/day: a full 5.3 at 744B/40B active is already expensive; Flash/Air lowers the bill.

Careful: “rumored 300B” is not an official sheet. Treat it as technical gossip, not a passport. The +75 tokenizer is still harder evidence than the size rumor.

Analogy, example, trap

Analogy. Coke and Coke Light. Same house recipe, less sugar. Taste betrays the house. Calories betray the SKU.

Example. iPhone and iPhone SE. Same store, different price. If you see an Apple logo on a cheap phone, it is not a Pixel.

Ox cut. The dossier’s leading thesis is “GLM-5.x multimodal family”. Flash/Air is the most likely last name, not the only one.

Lesson 17

What an agent is (why OpenCode gives it away)

A chat answers. An agent stays all afternoon touching your repo.

Agent = model + tools (terminal, browser, tests) in a loop. Read the error, change the file, run again. That eats tokens like crazy: the repo is pasted again and again.

Read all

That is why 97% cache matters: if the server remembers the repo, the loop does not break the bank. z.ai already did that. Ox Alpha does too.

OpenCode (terminal) and Hermes Agent / Claude Code / ZCode show up in the consumption top. ZCode is a z.ai product. Coincidence… or kitchen clue.

Analogy, example, trap

Analogy. Not a teacher explaining homework. An intern you leave the laptop with, who comes back at 2 a.m. with a PR. The intern drinks coffee (tokens) non-stop.

Example. A robot vacuum does not “talk about dust”. It loops. Ox Alpha is built to loop over code, not poetry.

Ox cut. The official sheet says it: coding, long agents, production. Not a café chatbot.

Lesson 18

Late theories (Composer, Cursor, Anomaly)

A new suspect is born every hour. Almost none survive the tokenizer.

On August 21 @TimMacc proposed Composer 3: a GLM fine-tune with Cursor compute. Fits “GLM underneath” and “lots of traffic”. Fits 1210 errors and 97% z.ai cache less well — unless Cursor is serving from there.

Read all

Anomaly / an OpenCode wrapper: collect data from free models. Possible, but 100T/day is a hyperscaler bill.

Golden rule: a new theory must explain the +75 AND the 1210 AND the video encoder. If it only explains the vibe, it is a meme.

Analogy, example, trap

Analogy. Someone says the tarped car is a Tesla tuned with a Toyota engine. Maybe. But the toll, the dashboard error and the radio are still Toyota.

Example. Reddit theories about “who the anonymous singer is” usually die when the contract drops. Until then, score the tests, not the hype.

Ox cut. The dossier leaves Composer/Cursor at ~8–12%. Alive, not leading.

Lesson 19

Why 2^20 and not a million

Computers count in twos. 1,024 is not 1,000. 1,048,576 is not a marketing million.

A bit is a switch: 0 or 1. With n switches you can represent 2^n things. 2^10 = 1,024 (the computer kilo). 2^20 = 1,048,576 (the mega).

Read all

GPUs and attention masks love sizes aligned to powers of two. That is why context windows are almost never a round 1,000,000.

Ox Alpha and GLM-5.2/5.3 publish exactly 1,048,576. Xiaomi also uses 1M, so this is not a passport. It does kill the lazy slogan.

Analogy, example, trap

Analogy. A 12-egg carton, not 10. The supermarket sells dozens because the box fits. Here the box is the GPU.

Example. A 64 GB stick does not have 64,000,000,000 round bytes. The number comes from powers of two, same as this context.

Ox cut. In the math lab you can move n and watch 2^n grow up to Ox's desk.

Formula

2^20 = 1,048,576   (Ox context = GLM-5.2/5.3 context)

An engineering coincidence, not marketing. Xiaomi uses it too: not enough alone.

Lesson 20

What an error log is (and why Chinese matters)

When a program breaks, it yells in the language of whoever programmed it.

A log is the server's diary. If you send an illegal parameter, the backend sometimes returns not just 400 but an internal paragraph.

Read all

On 21 August Josh rad sent invalid params to Ox Alpha and got Chinese logs plus strings that match GLM errors on ModelScope (China's Hugging Face).

A California lab does not leave the workshop manual in Chinese by accident. Composer 3 / Grok / Gemini look worse with this clue.

Analogy, example, trap

Analogy. You order a pizza with no logo. The complaint ticket is in Chinese and the code belongs to one shop. It is not Domino's in Ohio.

Example. If your taxi app fails and the error is errcode=1210 in Chinese, the backend is not Uber.

Ox cut. Adds to Pritish's 1210/1214. Two testers, same kitchen.

Lesson 21

1 GW: where 100T free tokens come from

Giving away tokens is not charity. It is a factory test with the lights paid.

Tom's Hardware / Bloomberg (21 Jul 2026): Z.ai stood up a 1-gigawatt data center on Chinese chips, several clusters of >10,000 chips. One GW powers ~750,000 homes.

Read all

100T tokens/day is an OpenCode capacity number. What OpenRouter showed on 21 Aug was ~2T cumulative: huge, not 100T. If Ox is Flash/Air, even cheaper per token.

Can Z.ai? is no longer a meme. Would they spend that on a stealth? yes: Pony Alpha already did.

Analogy, example, trap

Analogy. A factory opens on Saturday and gives away sneakers to see if the belt holds. It is not a corner shop.

Example. AAA game free weekend betas: they measure servers, not your happiness.

Ox cut. The they-don't-have-compute thesis dies with the 1 GW. The it's-Grok-because-expensive thesis too.

Lesson 22

How to think like an auditor (don't eat the hype)

One clue is gossip. Seven independent clues are a sketch. Zero official clues are a sketch, not a passport.

Rule 1: split the sheet (OpenRouter) from rumor (X). No URL, it does not enter.

Read all

Rule 2: a theory must explain tokenizer + errors + video. If it only explains the vibe, it is a meme.

Rule 3: Manifold 85% is a bet, not a paper. A 10-task quiz is not DeepSWE's 113. The viral 80% was n=10; Wenqi (DeepSWE's author) measured ~63%.

Rule 4: Mira's Inkling has been public since July and on 21 Aug it went free on OpenRouter with its name. Mixing Spanish mira (look) with Mira Murati is a gag.

Rule 5: the 9.9 ceiling needs our own tokenizer run or a lab statement. Until then: strong hypothesis.

Analogy, example, trap

Analogy. A detective does not close the case because a neighbor smells it. They close when there is a print, a shoe, a voice and a zip code. They still wait for DNA.

Example. Do not bet the rent because a tweet said confirmed. This dossier teaches you to ask for the token subtraction.

Ox cut. The Sources page is the chain of custody. Repeat the +75 subtraction before you tweet a verdict.